ECG Deep Learning Models Show Moderate Generalization Across Fatigue Conditions

Deep learning models designed to detect sleepiness from electrocardiogram recordings achieve moderate performance when evaluated across different participants and distinct fatigue-inducing conditions, according to a study by researchers at the University of Žilina in Slovakia. Investigators Ludmila Kralikova, Juraj Spanik, and Branko Babusiak assessed how well neural networks trained on prolonged wakefulness could identify subjective sleepiness during a separate driving task.

Fatigue and sleepiness can develop under varied behavioral and experimental circumstances, yet detection systems derived from electrocardiogram (ECG) data are frequently tested solely within the specific datasets used to construct them. The authors aimed to determine whether model performance holds up under participant-independent testing and when transferring between fundamentally different fatigue-induction paradigms.

To investigate this transferability, the researchers examined two ECG-derived analytical architectures developed using the DROZY database, where sleepiness was triggered through extended wakefulness. One approach utilized a one-dimensional convolutional neural network (1D CNN) based on six heart rate variability (HRV) metrics, while the second implemented a dual-input architecture combining convolutional and bidirectional long short-term memory units (CNN-BiLSTM) to analyze RR interval sequences alongside statistical HRV descriptors.

Internal evaluation within the DROZY database was performed using a five-fold subject-wise scheme across 25 recordings collected from 14 participants. In these internal tests, the HRV-based 1D CNN reached 60.0% accuracy, an F1-score of 61.5%, and an area under the receiver operating characteristic curve (ROC-AUC) of 0.593. The dual-input CNN-BiLSTM delivered higher internal baseline marks, posting 68.0% accuracy, a 73.3% F1-score, and a ROC-AUC of 0.753.

The trained networks were subsequently deployed without retraining or parameter adjustment to an independent dataset known as FatiQData, which tracked 20 participants performing a prolonged monotonous driving task. During this driving experiment, participant ratings on the Stanford Sleepiness Scale increased significantly from a baseline of 2.40 ± 0.82 to 4.15 ± 0.81.

In binary external classification on the driving cohort, the HRV-based 1D CNN attained 60.0% accuracy, an F1-score of 69.2%, and a ROC-AUC of 0.766, whereas the dual-input CNN-BiLSTM produced 65.0% accuracy, a 77.4% F1-score, and a ROC-AUC of 0.750. The authors noted that while cardiac RR dynamics and HRV metrics retain signals reflecting subjective sleepiness, participant-independent and cross-dataset detection capabilities remain moderate, underscoring the critical need for subject-wise validation and independent testing across heterogeneous experimental settings.

Subject of Research: Cross-dataset generalization of ECG-based deep learning models for sleepiness detection

Article Title: Generalizing ECG-based fatigue detection across fatigue-induction paradigms using deep learning

Article References: Kralikova, L., Spanik, J., & Babusiak, B. (2026). Generalizing ECG-based fatigue detection across fatigue-induction paradigms using deep learning. Frontiers in Digital Health, 8, Article 1911941. https://doi.org/10.3389/fdgth.2026.1911941

Image Credits: AI Generated

DOI: 10.3389/fdgth.2026.1911941

Keywords: electrocardiogram, fatigue detection, sleepiness, deep learning, heart rate variability, CNN-BiLSTM, DROZY database, FatiQData, Generalizing, ECG-based, fatigue, detection